Reject Inference Workflow

Fit on accepts, infer rejected outcomes, refit on the augmented set, and validate the lift

Reject Inference Workflow Fit on accepts, infer rejected outcomes, refit on the augmented set, and validate the lift 01 / Labeled Accepts 02 / Scorecard Model 03 / Reject Inference 04 / Augmentation 05 / Refit 06 / Validation EX / Fallback Reject-inference pipeline Reject inference If lift is negative Accepts · observed good/bad · Labeled Accepts › Reject-inference pipeline Accepts observed good/bad Base Model · fit on accepts · Scorecard Model › Reject-inference pipeline · biased sample Base Model fit on accepts biased sample Score Rejects · apply base model · Scorecard Model › Reject-inference pipeline Score Rejects apply base model Infer Outcomes · proxy good/bad · Reject Inference › Reject inference › Reject-inference pipeline Infer Outcomes proxy good/bad Combine Set · accepts + inferred · Augmentation › Reject-inference pipeline Combine Set accepts + inferred Refit Model · augmented data · Refit › Reject-inference pipeline Refit Model augmented data Validate Lift · held-out accepts · Validation › Reject-inference pipeline · gate Validate Lift held-out accepts gate Revert · keep base model · Fallback › If lift is negative › Reject-inference pipeline Revert keep base model Legend User UI Agent logic Policy Tool action Context / trace

One Honest Path

  • • The base model only learns from outcomes you truly observed
  • • Rejects are scored, inferred, then folded back in
  • • The augmented model is refit, not hand-tuned

Stop Conditions

  • • Negative held-out lift reverts to the base model
  • • Inference assumptions are documented before refit
  • • A degraded refit never ships silently

What To Retain

  • • The inference method and its missingness assumption
  • • Held-out comparison of base vs augmented model
  • • The weight or proxy assigned to each inferred row